Research map / Architectures and efficiency / Pruning and efficient tuning
WINS: Winograd Structured Pruning for Fast Winograd Convolution
Cheonjun Park, Hyun Jae Oh, Mincheol Park, Hyunchan Moon, Minsik Kim, Suhyun Kim, Myung Kuk Yoon, Won Woo Ro
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use l2-norm instead of l1-norm as representative value (ablation) | WINS · CIFAR-10 ResNet-32 (PR 90%) | Top-1 accuracy (%): 91.33 -> 91.84 | Improved |
| use Geometric Median instead of l2-norm (ablation) | WINS · CIFAR-10 ResNet-56 (PR 90%) | Top-1 accuracy (%): 92.08 -> 92.56 | Improved |
| use l2-norm instead of Geometric Median (ablation) | WINS · CIFAR-10 ResNet-20 (PR 80%) | Top-1 accuracy (%): 91.42 -> 91.51 | Improved |
| add load balancing across sub-GEMMs (WINS-B over WINS) (ablation) | WINS · ResNet-18 ImageNet | relative speedup: 1.15x at 30% PR to 1.43x at 70% PR | Improved |
| add balancing (WINS-B) to WINS (ablation) | WINS · ResNet-18 ImageNet (PR 80%) | Top-1 accuracy drop: 1.27% drop for WINS-B vs less than 1% for WINS | Hurt |
| add adaptive layer-wise selection (WINS-AB over WINS) (ablation) | WINS · ResNet-18 ImageNet (PR 80%) | speed: approximately 1.2 times faster | Improved |
| prune 1x1 convolutions with BCBP instead of FPGM (ablation) | WINS-AB · Res2Net-50 ImageNet (PR 85%/50%) | Top-1 accuracy (%): 75.23 -> 75.87 | Improved |
| prune 1x1 convolutions with BCBP instead of FPGM (ablation) | WINS-AB · Res2Net-50 ImageNet (PR 65%/30%) | Top-1 accuracy (%): 76.16 -> 76.32 | Improved |
| use WINS-AB instead of filter pruning | ResNet-18 · ImageNet | accuracy drop: FP exceeds 2.0% drop from 30% PR; WINS-AB does not exceed 0.6% at 80% PR | Improved |
| use WINS-AB instead of filter pruning | ResNet-18 and VGG-16 · ImageNet | speedup within 1.5% accuracy drop: 1.84x and 1.27x faster | Improved |
| use WINS-AB instead of Winograd weight pruning | VGG-16 · ImageNet (PR 70%) | accuracy: WWP 2.0% higher accuracy | Hurt |
| use WINS-AB instead of Winograd weight pruning | VGG-16 · ImageNet (PR 84%) | speedup: 2.01x faster than Winograd convolution | Improved |
Closest papers by meaning
- DEPrune: Depth-wise Separable Convolution Pruning for Maximizing GPU Parallelism (NeurIPS 2024 · similarity 0.88)
- OSSCAR: One-Shot Structured Pruning in Vision and Language Models with Combinatorial Optimization (ICML 2024 · similarity 0.85)
- REPrune: Channel Pruning via Kernel Representative Selection (AAAI 2024 · similarity 0.85)
- Data-Free Group-Wise Fully Quantized Winograd Convolution via Learnable Scales (CVPR 2025 · similarity 0.84)
- FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning (ICLR 2024 · similarity 0.84)
- ICP: Immediate Compensation Pruning for Mid-to-high Sparsity (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
